Learning task-specific subspaces via interventional post-training of speech foundation models